Carbon footprint of frontier AI models per query and per token

One query of 1210 tokens to a frontier model (class C, such as gpt-4, gpt-4-turbo, gpt-5) emits about 0.2 g CO₂e on the global-average grid, and a million tokens emit 0.147 kg CO₂e. Served from EU North - Sweden the same query emits 0.037 g, and from Asia Pacific - Singapore 0.21 g.

For comparison on the same grid, a class A (small) query emits 0.016 g and a class B (mid-size) query emits 0.053 g. Class C uses 0.26546404032860343 kWh of facility electricity per million tokens, and its uncertainty band around the class factor is about 67%.

EU North - Sweden gives the lowest total for class C, 0.0286 kg CO₂e per million tokens and Asia Pacific - Singapore the highest at 0.151 kg, a 5.3 times spread. The spread on electricity alone is 13 times.

A 1210-token query to a class C model consumes 0.54 mL of water in US South Central - Texas, the lowest-water region with a figure, and 3.5 mL in US West - Oregon, the highest. 14 of 18 regions have a water figure.

On the global-average grid, electricity is 87.4% of the class C total, hardware 2.6%, data-centre construction 1.2% and amortised training 8.8%. In EU North - Sweden electricity falls to 35.6% and in Asia Pacific - Singapore it rises to 87.8%. The conservative high case adds 58% to the class total.

The dataset's mapping table assigns 111 model name patterns from 24 providers to class C, and a star stands for any text. 13 of the 66 mapping rows for this class are flagged as reasoning models by default.

• OpenAI gpt-4, gpt-4-turbo, gpt-5, o1, o3, o3-pro, gpt-4-32k, gpt-5*-sol, gpt-6*-sol, gpt-6*-astra, gpt-5*-codex, gpt-6*-codex, gpt-5*-codex-max, gpt-6*-codex-max, gpt-5*-pro, gpt-6*-pro, gpt-5*-chat, gpt-6*-chat, gpt-chat-latest, chat-latest, gpt-6, o3-deep-research, o4-mini-deep-research
• Anthropic claude-opus-*, claude-3-opus*, claude-fable-*, claude-*-fable, claude-mythos-*, claude-*-mythos, claude-*-opus, claude-2, claude-v2, claude-1, claude-v1
• Alibaba qwen3-235b-a22b, qwen*-max, qwen3-vl-235b-a22b, qwen3-omni-235b-a22b, qwen*-2.4t-a95b, qwen*-397b-a17b, qwen*-72b, qwen*-110b, qwen*-180b, qwen
• DeepSeek deepseek-v3, deepseek-chat, deepseek-r1, deepseek-reasoner, deepseek-v4, deepseek-v4*-pro, deepseek-v4*-flash, deepseek-flash, deepseek-v2, deepseek-coder
• Meta llama-3.1-405b, llama-4-maverick, llama-4-behemoth, llama3-2-90b, llama3-1-405b, llama3-405b, llama4-maverick, muse-spark, muse-glimmer, muse-code
• Cohere command-r-plus, command-a, command, north-small
• MiniMax minimax-m, minimax-h, minimax-text, abab
• Mistral mistral-large, mistral-medium-3, mistral-large-3, pixtral-large
• Moonshot AI kimi-k2, kimi-k3, kimi-k1, moonshot-v1
• Tencent hy4, hunyuan-large, hy3, hunyuan
• Xiaomi mimo-v2, mimo-*-flash, mimo-*-pro, mimo
• AI21 jamba-large, jamba-1.5-large, jamba-instruct
• Amazon nova-premier, nova-*-premier, titan-text-premier
• NVIDIA nemotron-*-ultra, nemotron-ultra, nemotron-4-340b
• Z.ai (Zhipu) glm-5*-flash, glm-, zai-glm
• Meta/Alibaba llama-3.2-90b*, qwen2.5-72b
• Perplexity sonar-reasoning*, sonar-deep-research
• AionLabs aion-
• any *
• Google gemini-ultra
• Thinking Machines inkling
• Typesafe jev-
• Writer palmyra
• xAI grok-*

• Class A (small)
• Class B (mid-size)
• Class C (frontier)

Figures per query are one plain exchange, before the Tier 2b reasoning uplift. Electricity term. GPU energy per token from the ML.ENERGY Leaderboard v3, a facility overhead of 1.20, and grid carbon intensity from EPA eGRID 2023 and Ember 2023. Embodied hardware, data-centre construction and amortised training terms are documented on the methodology page. Water combines on-site cooling water (Li et al. 2025) with the water embedded in grid electricity (Reig et al. 2020). The modelled range of the models inside a class, around the class value, is −41% to +48% for class A, −39% to +24% for class B and −67% to +55% for class C.

Dataset: AI Inference Emission and Resource Factors for Corporate GHG Inventories (SOMA AI Emission Factors (SAEF)), Llopis-Garcia, G., SOMA AI, 19 October 2026, CC BY 4.0, concept DOI 10.5281/zenodo.20443585, which resolves to the latest version (this version 10.5281/zenodo.23371447, earlier version 2 values at 10.5281/zenodo.22767475). Method: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, arXiv:2606.10660, under peer review.

Partners can fetch any row from the AI factors API, and the whole table is a CSV download.

The full table for every class and region is on the SOMA AI Emission Factors (SAEF) page.

• Carbon footprint of small AI models (class A)
• Carbon footprint of mid-size AI models (class B)
• How the factors are calculated
• How to account for AI usage in a Scope 3 inventory, and when CSRD applies